Diffusion Jev
Independent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.
Overview
Diffusion Jev is an independent implementation of Jev-style typed decisions on Google's DiffusionGemma, served with SGLang, with no fine-tuning. One model powers three demos in a web playground: draw a doodle and it picks from 16 labels with scores for each; choose a flower photo and it picks among five classes from the pixels; write a message and it returns an emoji choice, a yes/no answer and an intensity score.
It supports Choice, Score and Noul for both text and images and publishes evaluation artifacts, including a three-model comparison. According to the list summaries, option scores come from self-conditioned denoising logits and are explicitly not calibrated correctness probabilities. The public app's GPU sleeps between visits, so the first prediction can take minutes. It is a research demo of reading decisions from a diffusion language model.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of the 16 labels does this doodle show? | airplane · apple · bicycle · cat · clock · fish · pizza · umbrella · dog · car · house · tree · sun · star · cup · sailboat |
| Choice | Which flower is in this photo? | daisy · dandelion · rose · sunflower · tulip |
| Score | How intense is the sentiment of this message? | 0 · 1 · 2 · 3 |
How it works
- User draws, uploads or types input
- Server poses typed questions to DiffusionGemma
- Option scores are read from denoising logits
- Playground shows the pick and all candidate scores
Strengths
- Works on images and text without fine-tuning
- Public evaluation artifacts
- Interactive playground
Caveats
- Scores are not calibrated probabilities
- Cold starts of several minutes on the public demo
- Needs a large GPU to self-host
Getting started
Try https://diffusion-jev-sglang.vercel.app or follow docs/diffusiongemma.md to self-host.At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- SGLang server + web playground
- Has code
- yes
Use cases and integrations
How curated lists describe it
Visual classification: independent Jev-style DiffusionGemma/SGLang server that selects doodle and flower labels from image pixels with typed Choice questions and displays candidate scores in a drawing playground, with public evaluation artifacts and uncalibrated probabilities.
yibie/awesome-jevIndependent DiffusionGemma/SGLang implementation of Choice, Score and Noul for text and images, with a doodle-classification playground and public evaluation artifacts; option scores come from self-conditioned denoising logits and are not calibrated correctness probabilities.
cobanov/awesome-jev
Frequently asked questions
Is this TypeSafe's Jev?
No, it is an independent Jev-style server on DiffusionGemma.
Are the scores calibrated?
No, the summaries note they come from denoising logits and are not calibrated.
Why is the demo slow at first?
The GPU sleeps between visits, so the first prediction includes a cold start.
Related projects
Jev-style choice-plus-probabilities from GLM-5.3-Flash in one forward pass, in Edgeless Systems' confidential-computing API.
ServicebetaPythonOpen Jev-style text-and-image decision models built by a self-improving AI research loop, with an effort setting that picks depth.
ServiceresearchPythonSemIf (formerly OpenJev): reads typed option probabilities directly from open models like Qwen3.5-4B, including image inputs, on a home GPU.
ModelresearchPython- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPython